I've spent months building a small team of specialized AI agents inside Claude: one drafts this newsletter, one tracks invoices and client health, one does market research, one runs a trading system's operations. Each one runs on its own playbook, and each one hands anything irreversible back to me before it happens.
Turns out I was hand-building a prototype of a product Anthropic already shipped.
Tool of the Week: Claude for Small Business
Claude for Small Business, live since May 13, packages that same idea for anyone running a small company or solo practice: 15 ready-to-run agentic workflows and 15 task-specific skills, covering the parts of running a business that everyone hates: payroll planning, month-end close, invoice chasing, lead triage, contract review, campaign creation, cash-flow monitoring, and a daily "business pulse" report. It connects to the tools you probably already use: QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace, Microsoft 365.
How it actually works: you turn it on inside Claude Cowork, connect the accounts it needs, and pick a workflow. Claude does the work (drafts the reminder emails, reconciles the books, triages the leads, flags the contract clauses) and you approve before anything actually sends, posts, or pays. That approval gate is the whole point. It's the same rule I use by hand: automate the prep and the drafting, keep a human on the button that can't be undone.
What I'd actually do with this, in order:
Don't turn on all 15 at once. That's the fastest way to lose trust in the whole system: one workflow does something slightly wrong in week one and you shut the entire thing off. Pick one.
Pick the task that costs you the most time and annoys you the most, not the flashiest one. For most small business owners that's invoice chasing or lead triage, not campaign creation. Boring and expensive beats interesting and occasional.
Run it for two weeks before adding a second workflow. Watch what it drafts. Correct it. Once it just works, add the next one.
Skip what doesn't apply to you. If you have zero employees, ignore the HR onboarding workflow. Scan the list for the two or three that map to a task you actually have.
Who this is for: anyone already talking to Claude for drafting and research who's never had it actually touch the back-office grind: the invoicing, the reconciling, the lead sorting. That's the gap this closes. You're not getting a fully autonomous employee. You're getting a very fast assistant who does the prep work and waits for your sign-off. If you've been burned by "autonomous AI agent" hype before, that should sound like a feature, not a limitation.
Quick Hits
An AI model may have just done graduate-level pure math, with proofs mathematicians vouch for. On August 1, OpenAI said an internal version of its next model, Astra, solved ten previously open problems in mathematics and theoretical computer science, including proving the existence of "non-sofic groups" and new sphere-packing bounds, and published the results as formal, machine-checkable proofs. Fields Medal winner Timothy Gowers said he'd recommend one of the proofs for a top journal without changes. Why it matters: you don't need to do anything about this today, but it's a real signal that frontier reasoning is still climbing fast, into territory that used to be considered decades from automatable. The gap between "good enough to draft an email" and "good enough for genuinely novel expert work" is closing quicker than most roadmaps assume.
Running AI on your business just got cheaper again. On July 30, OpenAI cut prices on its GPT-5.6 Luna model by roughly 80%, to $0.20 per million input tokens and $1.20 per million output tokens. DeepSeek's competing V4 Flash model already undercuts that, at roughly $0.14 in / $0.28 out, with automatic caching pushing repeat-input costs even lower. Why it matters: if you shelved an automation idea earlier this year because the AI cost didn't pencil out, redo that math. The per-task cost of running AI keeps falling. Ideas that didn't work in the spring might work now.
The tools you already use got faster this month, and you didn't have to do anything. OpenAI said in August that improvements to how its models generate text made GPT-5.6 more than 15% more efficient, and rolled the gains out automatically across the API, Codex, and ChatGPT. Why it matters: not every AI improvement requires switching tools or learning something new. Some of it just shows up as a faster reply and, eventually, a smaller bill. Worth remembering next time a familiar tool "feels different." Sometimes it's not you, it's a quiet upgrade underneath.
Prompt of the Week: The First Workflow Prompt
Before turning on any pre-built AI workflow (Claude for Small Business or otherwise), use this to pick which ONE to try first:
Act as an operations consultant helping me choose ONE recurring business
task to hand to a pre-built AI workflow, not several, just one, as a pilot.
I'll list the repetitive tasks that eat my time each week or month. For
each one, tell me:
- Roughly how many hours a month this actually costs me
- How reversible a mistake would be if the AI got it wrong (a typo in a
draft vs. money sent to the wrong place)
- Whether it needs my judgment or relationships, or is mostly mechanical:
the same steps every time
Then rank them and tell me which ONE to automate first, favoring high
time-cost, low-risk-if-wrong, and mostly-mechanical tasks, and which
ones I should NOT hand to AI yet, and why.
Here are my recurring tasks:
[list the repetitive things you do in your business each week or month]Most people over-automate on day one and regret it. Pick the boring, expensive-in-time, low-risk task first. Prove it works. Then add a second.
Like what you're reading? Forward it to someone who'd get value from it. And if you're curious what AI could actually do inside your business, book a free 15-minute audit. No pitch, just a look at where you're leaving time on the table.